A scene-acoustic-fingerprint-based medical device alarm sound identification method and system

Abstract

This invention discloses a method and system for identifying alarm sounds from medical devices based on scene-based voiceprints. Addressing the technical problem of insufficient anti-interference capability in alarm sound recognition in noisy ward environments, the method simultaneously extracts scene voiceprints and foreground sound features from the audio segment to be identified. Scene voiceprints are then classified to determine their scene category. A noise level estimate is obtained by querying a scene category-noise level mapping table based on the scene category. The comparison between this noise level estimate and a preset noise threshold is used as a path selection condition. If the noise level is below the threshold, a digital signal processing waveform template matching path is used; if it is above the threshold, a preset scene-alarm joint mapping table matching path is used. The identification target is limited to a finite set of pre-registered target alarms in the hospital. This invention solves the false alarm problem caused by waveform superposition in noisy environments using traditional noise reduction-then-identification schemes, achieving low false alarm and low missed alarm identification for alarm sounds from medical devices without protocol interfaces.

Description

Technical Field

[0001] This invention relates to a method and system for identifying alarm sounds of medical devices based on scene voiceprints, belonging to the field of intelligent sensing and acoustic signal processing technology for medical devices. Background Technology

[0002] In clinical settings such as hospital wards, intensive care units, and operating rooms, medical equipment such as infusion pumps and monitors use audible alarms to alert clinical nurses to equipment malfunctions or changes in patient condition. In modern large hospitals, dozens to hundreds of medical devices may operate simultaneously in a single ward, generating as many as thirty to forty different alarm sounds. Existing research and clinical experience indicate that clinical nurses can typically only effectively identify three to four types of alarm sounds, with the remaining alarm sounds often ignored or misinterpreted, posing potential clinical safety hazards.

[0003] In the existing technology, the main solutions to the above problems include the following categories:

[0004] The first type of solution is device data acquisition based on protocol interfaces. This type of solution directly reads the alarm status words generated by the medical device through the digital communication interface. When the device has an open protocol, it can achieve accurate acquisition of alarm status. The inherent drawback of this type of solution is that many micro-infusion pumps, older models of ventilators, and some imported monitoring devices that are widely used in clinical practice do not provide accessible digital protocol interfaces, resulting in an inability to cover all device alarm events.

[0005] The second type of scheme is waveform template matching based on digital signal processing. This scheme performs analog-to-digital conversion on the input audio and then identifies specific alarm tones by shape matching with a preset waveform template. It can achieve near-perfect accuracy in quiet environments. The main drawback of this scheme is that when multiple background noises are present and the sound sources come from different directions, the target alarm tone waveform and the background acoustic signal linearly superimpose. The template matching loses its ability to distinguish the superimposed waveform, resulting in numerous false alarms.

[0006] The third type of approach is based on cascaded processing of noise reduction followed by identification. This involves using signal processing techniques such as spectral subtraction and Wiener filtering to preprocess the noisy frequency signal for noise reduction, and then performing template matching or classification on the denoised signal. The fundamental flaw of this approach is that it implicitly assumes that background noise and the target signal are separable in the time-frequency domain. However, the background acoustic signal in a hospital ward contains many non-stationary components, such as conversations, footsteps, equipment fans, and similar alarm sounds from other equipment. After noise reduction, the characteristics of the target signal are simultaneously weakened, leading to an increased false negative rate.

[0007] The fourth type of solution is a recognition scheme based on a general sound event detection model, which typically uses convolutional neural networks or recurrent neural networks to classify sound events in an open set. The main shortcomings of this type of solution in medical clinical scenarios are: the limited number of target types to be identified (usually three to four) and the highly specialized acoustic environment (fixed ward acoustic scenario); the general open set model does not utilize the aforementioned prior knowledge, making it difficult to reduce the false alarm rate on small target set tasks; furthermore, the general model requires a large deployment scale, making it difficult to deploy locally or at the edge of hospitals to meet the compliance requirements of keeping medical data within the hospital premises.

[0008] The fifth type of solution is a system based on alarm hierarchy management. This type of solution classifies alarm events into high, medium, and low levels and prioritizes pushing high-level alarms. Its limitation is that it only solves the management and distribution problems after the alarm event is generated, and does not solve the problem of how to reliably identify the target alarm sound in a noisy ward environment at the acoustic level.

[0009] The sixth type of scheme is a sound event detection scheme based on acoustic scene cascade recognition. This involves first performing scene category recognition on the input audio, and then selecting a corresponding classifier from several pre-trained classifiers based on the recognized scene category for sound event recognition (e.g., the technical solution disclosed in Chinese invention patent application CN104795064A). This type of scheme has the following shortcomings: the scene category is only used to "select the most suitable recognition model," and the output space remains an open set, meaning that non-target sound events may be misclassified as target events in the recognizer's output space; this type of scheme also does not limit the recognition target to a finite set of pre-registered target alarms in medical facilities. When multiple alarm sounds of the same brand are simultaneously mixed into a ward, alarm events belonging to non-target devices may still be output. The aforementioned scenario-driven multi-model selection approach also has parallel variations in Chinese patent literature for human voice speech recognition, such as the technical solution disclosed in Chinese invention patent application CN116364061A. This solution identifies the scene category using a scene classifier and then selects the corresponding model from multiple pre-trained speech recognition models to recognize human voice speech. This parallel variation differs from the technical solution based on CN104795064A in the recognition target (the former focuses on human voice speech, while the latter focuses on sound event detection). However, both are open-set single-path architectures, neither limits the recognition target to a pre-registered finite set of alarm targets, and neither involves mutually exclusive splitting between the neural network path and the digital signal processing waveform template matching path based on noise level thresholds. Therefore, they are classified as having the same deficiency.

[0010] The seventh type of scheme is an audio scene classification scheme based on neural networks. This involves using a two-stage neural network to classify audio scenes and using the scene classification output as an auxiliary signal for downstream signal processing (e.g., the audio scene classification scheme for hearing aids disclosed in US patent application US20210174824A1). The shortcomings of this type of scheme are: the scene classification output only serves as an auxiliary clue for adjusting downstream processing parameters and does not constitute a constraint on the legitimacy of the identified target; the application domain is subjective auditory optimization (generating hearing aid stimulus signals), not an alarm tone recognition task aimed at determining the existence of objective events.

[0011] The eighth type of solution is an acoustic patient monitoring solution based on microphone arrays. This involves capturing and directly classifying acoustic events emitted by the patient's body, such as coughing, snoring, groaning, and falls / collisions, using a microphone array (e.g., the technical solutions disclosed in US Patent No. US20110087079A1 and Chinese Invention Patent Application No. CN102088911A). The limitations of this type of solution in identifying medical device alarm sounds are: the object of identification is the patient's bodily acoustic events (human sounds), not the alarm acoustic events generated by medical devices (machine sounds). These two types of acoustic signals differ fundamentally in their spectral structure, temporal characteristics, and semantic intent. Furthermore, this type of solution directly classifies acoustic features without a scene-conditioning mechanism, and cannot solve the problem of legality filtering for multiple device alarm sounds in noisy ward environments.

[0012] The ninth type of scheme is a single-path alarm tone recognition scheme based on target location cutting and correlation coefficient calculation. This involves preprocessing the sampled alarm audio to obtain a first audio, then cutting out a second audio from the first audio based on the condition that "the similarity between the first audio and the standard audio meets a threshold," and determining whether it is identified as the target alarm tone based on the correlation coefficient between the second audio and the standard audio (e.g., the technical solution disclosed in Chinese invention patent application CN117095698A). This type of scheme has several shortcomings: the determination is based solely on the similarity and correlation coefficient between a single audio stream and a single standard audio, lacking explicit modeling of the acoustic scene and a scene-conditional diversion mechanism; it lacks any conditional path switching oriented towards noise level ranges; the recognition object is the shape matching between a single alarm tone and a single standard audio, without involving the constraint of a closed set whitelist that limits the recognition target to a finite pre-registered set of target alarms; when the tail sounds of similar alarm tones from other devices are simultaneously mixed in the ward environment, the position of the maximum correlation coefficient may be hijacked by the alarm tone of a non-target device, leading to false alarms.

[0013] The tenth category is a neural network model selection scheme driven by acoustic scene recognition. This involves identifying the acoustic scene through an acoustic scene recognition module and autonomously selecting different neural network models from the neural network noise reduction or processing module to perform downstream signal processing based on the identified scene category (e.g., the environment-adaptive neural network noise reduction scheme for digital hearing aids disclosed in Chinese invention patent application CN109859767A). The shortcomings of this type of scheme are: the scene recognition result is only used as a "selection signal" for the downstream neural network model, meaning its purpose is to decide which model to use from several pre-trained noise reduction or processing models, rather than determining whether the input acoustic event belongs to a pre-registered set of legitimate targets; the application domain is subjective hearing optimization for hearing aids (improving speech intelligibility) rather than objective event existence determination; its output space is the noise-reduced audio signal or subjective hearing signal, which does not constitute a precise classification of alarm categories; and it does not involve mutually exclusive branching between the neural network path and the DSP waveform template matching path based on noise level thresholds.

[0014] In summary, existing technologies either lack the ability to conditionalize scenarios (Solutions of Categories 2 / 3 / 4 / 5 / 8 / 9), or although they have scenario cascading capabilities, the scenarios are only used for model selection and do not constitute a legal constraint on the recognition target (Solutions of Categories 6 / 10), or scenario classification is only used as a downstream auxiliary signal rather than an access control (Solution of Category 7), or the recognition object is human voice rather than machine voice (Solution of Category 8), or they do not impose closed set constraints on the recognition target's pre-registration of a finite target alarm set (Solutions of Categories 6 / 7 / 8 / 9 / 10), or they do not perform mutual exclusion splitting of neural network paths and DSP waveform template matching paths according to noise level thresholds (Solutions of Categories 1 to 10). Summary of the Invention

[0015] The purpose of this invention is to provide a method and system for identifying medical device alarm sounds that can be stably identified in noisy hospital ward acoustic environments, while also achieving high accuracy in identifying small target sets and being locally deployable.

[0016] To achieve the above objectives, the first aspect of the present invention discloses a method for recognizing alarm sounds of medical devices based on scene voiceprints, characterized by comprising the following steps: Step 1: Receive the audio segment to be identified from the audio acquisition terminal deployed in the medical facility. The audio segment to be identified is obtained by segmenting the continuous audio stream through a fixed-duration sliding window. Adjacent sliding windows are allowed to overlap. The last segment with a length less than the preset window duration is filled with zero value to the standard length and then processed. Step 2: If the overall energy of the audio segment to be identified is lower than the preset silence threshold, it is directly determined as a non-alarm event, and the process returns to Step 1 to process the next sliding window. Otherwise, foreground sound feature extraction and scene voiceprint extraction are performed simultaneously on the audio segment to be identified to obtain foreground sound features and scene voiceprint; Step 3: Perform scene classification on the scene voiceprint, calculate the similarity between the scene voiceprint and several pre-constructed scene category prototype vectors, and determine the scene category corresponding to the scene category to which the scene voiceprint belongs by the scene category with the highest similarity and exceeding the preset scene classification threshold. If the similarity of all scene category prototype vectors is lower than the preset scene classification threshold, the current scene is determined to be an unknown scene and processed according to the preset fallback strategy. The preset fallback strategy includes either classifying the current audio segment into the preset default noise level and continuing to execute step 4, or determining the current audio segment as a non-alarm event and returning to step 1 to process the next sliding window. Step 4: Based on the scene category to which the scene voiceprint belongs or the unknown scene fallback strategy, query the pre-built scene category-noise level mapping table to obtain the current noise level estimate. Step 5: Determine the audio segment to be identified according to the following conditions and rules: When the noise level estimate is less than or equal to a preset noise threshold, the foreground sound features are matched with each template in the preset waveform template library. If the matching similarity of at least one template exceeds the preset template matching threshold, the target alarm event corresponding to the template with the highest matching similarity is taken as the judgment result. If multiple templates have the same highest matching similarity, the target alarm event corresponding to the template with the highest priority is taken as the judgment result according to the preset alarm priority. Otherwise, it is determined as a non-alarm event. When the noise level estimate is greater than the preset noise threshold, the foreground sound features and the scene voiceprint are constructed as binary feature inputs and jointly matched with the preset scene-alarm joint mapping table. If the joint matching similarity of at least one entry exceeds the preset joint matching threshold, the target alarm event corresponding to the entry with the highest joint matching similarity is taken as the judgment result; otherwise, it is judged as a non-alarm event. The identification targets of the preset waveform template library and the preset scene-alarm joint mapping table are limited to the target alarm set pre-registered by the hospital, and the target alarm set includes at least one target alarm sound type; Step 6: Output the results of the valid alarm events as alarm type labels, and convert them into voice prompts or text overlay information to push to the clinical terminal.

[0017] Preferably, in step 2, the foreground sound features include at least one of normalized spectral energy distribution, zero-crossing rate, temporal envelope peak value, and fundamental frequency estimate; the scene sound signature includes a statistically aggregated feature vector based on Mel frequency cepstral coefficients in the sliding window time dimension.

[0018] Preferably, in step 4, the scene category-noise level mapping table contains at least two scene categories, wherein at least one scene category corresponds to a noise level higher than the preset noise threshold, and at least one scene category corresponds to a noise level lower than or equal to the preset noise threshold; the noise level estimate corresponding to each scene category in the scene category-noise level mapping table is obtained by normalizing the signal-to-noise ratio statistical mean of the training samples of the corresponding scene category.

[0019] Preferably, the method for constructing the preset scenario-alarm joint mapping table includes: Step 501: Collect mixed acoustic samples of multiple scenes in the actual acoustic environment of the hospital ward, and label each sample with the scene category and whether it contains the target alarm sound and its type. The types of target alarm sounds are customized by the hospital clinical nursing team according to the actual needs of the department, and the number of types is limited to one to ten. Step 502: Using a pre-trained audio representation model as the teacher model, extract scene voiceprint vectors and foreground sound feature vectors from the labeled samples to form a triplet training sample set including the scene voiceprint vectors, foreground sound feature vectors, and alarm type labels; wherein, the scene voiceprint vectors are extracted from the output of the shallower intermediate layer of the teacher model, and the foreground sound feature vectors are extracted from the output of the deeper intermediate layer of the teacher model or the layer before the event classification head, so that the scene voiceprint vectors represent the overall statistical properties of the acoustic scene and the foreground sound feature vectors represent the local structural properties of the sound event; Step 503: Use the triplet training sample set to train a lightweight student model using the knowledge distillation method. The training objectives of the lightweight student model include the distillation loss between the output feature representation of the teacher model and the classification cross-entropy loss for alarm type labels. Step 504: After quantization and compression, the trained lightweight student model is deployed on the hospital's local edge computing device as an executable implementation of the scene-alarm joint mapping table, which is called by the scene-based voiceprint-based medical device alarm sound recognition method when the joint mapping path is selected.

[0020] Preferably, in step 501, the acquisition scenarios of the multi-scenario mixed acoustic samples include at least three scenarios: a quiet general ward, a general ward where people are talking, and an intensive care unit ward where multiple devices are running simultaneously; in step 503, the number of parameters of the lightweight student model does not exceed 100 million and does not exceed one-tenth of the number of parameters of the teacher model.

[0021] Preferably, the target alarm set includes one to ten alarm tone types. The target alarm set is customized by hospital administrators or clinical nurses according to the actual clinical needs of the department, and can be reconfigured locally in the hospital through a configuration interface.

[0022] Preferably, the method further includes a step of signal enhancement processing of the foreground sound features. The signal enhancement processing includes energy normalization and main peak gain of the foreground sound features based on a preset alarm audio spectrum peak template. The signal enhancement processing is performed before the template matching step described in step 5. The activation condition is that the noise level estimate is located in a preset intermediate interval, and the upper limit of the preset intermediate interval does not exceed the preset noise threshold.

[0023] Preferably, the method further includes a step of suppressing the background acoustic components represented by the scene soundprint. The suppression process includes subtracting steady-state background spectral components that are unrelated to the peak position of the target alarm audio spectrum from the scene soundprint. The suppression process is performed before the joint matching step described in step 5, and its activation condition is that the noise level estimate is higher than the preset noise threshold.

[0024] Preferably, in step 1, the audio acquisition terminal is a microphone or microphone array deployed at the bedside in the ward, operating room, or corridor. The medical device alarm sound recognition method and the medical device data acquisition system based on the protocol interface operate in parallel, and the alarm recognition result and the protocol acquisition result are sent together to the hospital clinical information system.

[0025] The second aspect of the technical solution of this invention discloses a medical device alarm sound recognition system based on scene voiceprints, characterized in that it includes: The audio receiving module is used to receive audio segments to be identified from audio acquisition terminals deployed in medical facilities. The audio segments to be identified are obtained by segmenting a continuous audio stream through a fixed-duration sliding window. Adjacent sliding windows are allowed to overlap. Segments that are less than the preset window duration at the end are filled with zero values ​​to the standard length and then used as the audio segments to be identified. The feature extraction module is used to simultaneously perform foreground sound feature extraction and scene voiceprint extraction on the audio segment to be identified, and output the foreground sound features and scene voiceprint; and when the overall energy of the audio segment to be identified is lower than the preset silence threshold, directly determine that the current audio segment is a non-alarm event and use the next sliding window as a new audio segment to be identified; The scene classification module is used to calculate the similarity between the scene voiceprint and several pre-constructed scene category prototype vectors respectively, and determine the scene category corresponding to the scene category to which the scene voiceprint belongs by the scene category corresponding to the scene category prototype vector with the highest similarity and exceeding the preset scene classification threshold; and determine the current scene as an unknown scene and process it according to the preset fallback strategy when the similarity of all scene category prototype vectors is lower than the preset scene classification threshold. The path arbitration module is used to query a pre-built scene category-noise level mapping table to obtain the current noise level estimate based on the scene category to which the scene voiceprint belongs or the unknown scene fallback strategy, and select the subsequent determination path based on the comparison result of the noise level estimate and the preset noise threshold. The DSP matching module is used to perform template matching between the foreground sound features and each template in the preset waveform template library when the path arbitration module selects a digital signal processing path. If the matching similarity of at least one template exceeds the preset template matching threshold, the target alarm event corresponding to the template with the highest matching similarity is taken as the judgment result. If the matching similarity of multiple templates is tied for the highest, the target alarm event corresponding to the template with the highest priority is taken as the judgment result according to the preset alarm priority. Otherwise, it is determined as a non-alarm event. The joint matching module is used to construct the foreground sound features and the scene voiceprint as binary feature inputs when the path arbitration module selects the joint mapping path, and perform joint matching with the preset scene-alarm joint mapping table. Based on the comparison result of the joint matching similarity and the preset joint matching threshold, it is determined whether it is the target alarm event corresponding to the corresponding entry. The alarm output module is used to output the results of valid alarm events as alarm type labels, and convert them into voice prompts or text overlay information to be pushed to the clinical terminal. The storage module is used to store and maintain the preset waveform template library, the preset scene-alarm joint mapping table, and the preset scene category-noise level mapping table. The identification targets of the preset waveform template library and the preset scene-alarm joint mapping table are limited to a target alarm set consisting of one to ten alarm sound types pre-registered by the hospital.

[0026] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention reduces the impact of waveform superposition on the recognition results. The cascaded structure of noise reduction followed by recognition implicitly assumes that the target signal and noise are separable in the time-frequency domain. When the background acoustic signal is a non-stationary multi-source mixture, this assumption fails, leading to recognition errors. This invention addresses this problem by handling it separately according to noise level: when the noise level corresponding to the category to which the scene soundprint belongs is lower than a preset threshold, the DSP template matching path is activated. Since the background noise energy is lower than the threshold ratio of the target alarm sound energy at this time, the spectral distortion caused by waveform superposition is limited to the preset template matching similarity tolerance range; when the noise level is higher than the preset threshold, the scene-alarm joint mapping path is activated. This path uses the binary features composed of scene soundprint and foreground sound features as the judgment input. The judgment process does not rely on direct comparison at the waveform level, thus bypassing the feature distortion caused by waveform superposition. The activation conditions of the above two paths are strictly mutually exclusive through noise level lookup table and threshold comparison. The decoupling logic can be repeatedly verified by configuring the noise threshold.

[0027] (2) Regarding the false alarm rate, this invention strictly limits both the template entry set of the preset waveform template library and the entry set of the preset scene-alarm joint mapping table to the target alarm set pre-registered by the hospital. In this case, for any scene category c, the set of output target alarm sounds A(c) of the joint mapping table in that scene is a finite closed set. If the binary input of the foreground sound features and the scene soundprint is mapped to a candidate category outside of A(c), it will be filtered out in the joint matching similarity threshold determination stage and will not constitute the final output. Compared with the open-set general sound event detection model, the entry space of the joint mapping table is smaller in magnitude, and each entry can obtain a higher sample coverage density under the same scale of labeled samples, which is beneficial to improving the confidence of the determination. Compared to correlation coefficient matching schemes based on a single audio stream and a single standard audio (such as the ninth prior art scheme), this invention changes the decision boundary for "whether to output as a target alarm event" from "the similarity with a standard audio exceeds a threshold" to "hitting the set A(c) that is pre-registered as a legal output under scene category c". That is, the decision boundary is raised from single-point similarity comparison to the determination of the subordinate relationship of a limited set of targets. When multiple alarm sounds of the same brand are mixed into the ward at the same time, even if the maximum value of the correlation coefficient of the tail sound is higher than the threshold, if the device to which the tail sound belongs has not been pre-registered as a target alarm source by the hospital, its corresponding target alarm sound category is not within A(c), so it is still excluded from the final output.

[0028] (3) Edge deployability is supported by the matching relationship between the student model parameter size and the computing power specifications of the ARM architecture inference box. The student model is deployed on the local inference box after quantization and compression. The upper bound of the model parameter size is derived and set according to the available memory and floating-point computing power of the local inference box (typically, when the upper bound of the parameter size does not exceed 100 million, it can run on an ARM A55 or A76 processor with a main frequency of 2GHz; preferably, the parameter size is controlled at the level of tens of millions, and forward inference can be completed within a 1.5-second sliding window). During the training phase, the student model is guided to learn the teacher model's representation of the scene-alarm joint features through the distillation loss between the intermediate layer representation of the teacher model and the teacher model. The corresponding weight coefficients have been given recommended value ranges in the specification for easy reproduction. Local deployment eliminates the need to upload patient audio to the hospital's external network, which is in line with the Level 3 Information Security Protection Requirements for the export of medical data.

[0029] (4) When used in conjunction with a protocol-interface-based device data acquisition scheme, this invention can cover a large number of medical devices such as micro-infusion pumps and older ventilators without protocol interfaces, forming a complete device sensing system that combines protocol acquisition and acoustic recognition. The two schemes complement each other in terms of device coverage. At the preprocessing level, the activation conditions for signal enhancement processing and background suppression processing are strictly mutually exclusive by noise level intervals. At most one of the two processing branches can be activated at any given noise level, avoiding the problem of inconsistent results caused by different processing orders when multiple preprocessing methods are activated simultaneously in the prior art. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 A schematic diagram of a dual-path decision structure for scene-based voiceprint conditionalization; Figure 3 A schematic diagram illustrating the process of constructing a scene-alarm joint mapping table; Figure 4 This is a schematic diagram of the module composition of the system of the present invention. Detailed Implementation

[0031] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0032] The core concept of this invention lies in the following: instead of employing a cascaded structure of noise reduction followed by recognition, it simultaneously extracts scene fingerprints that characterize the acoustic scene and foreground sound features that characterize suspected alarm sound components from the audio segment to be identified. The subsequent decision path is determined by the category to which the scene fingerprint belongs—when the noise level corresponding to the category is below a preset threshold, a digital signal processing waveform template matching path is used for determination; when the noise level corresponding to the category is above the preset threshold, a preset scene-alarm joint mapping table matching path is used for determination; and the identification target is limited to a pre-registered, bounded target alarm set of the hospital. With this arrangement, the false alarm problem caused by waveform superposition is transformed into a scene-conditional joint feature recognition problem.

[0033] In this embodiment of the invention, the "scene soundprint" specifically refers to a feature vector characterizing the overall statistical attributes of the acoustic scene in which the audio segment is located, which is different from the audio fingerprint used to uniquely identify a specific audio segment in the field of audio fingerprint recognition (such as hash-based song recognition)—the former characterizes the scene category, while the latter characterizes the individual audio segment. The "foreground sound feature" specifically refers to a feature vector characterizing the suspected alarm sound target component in the audio segment, which is different from the mixed audio features used by general sound event detection models in the prior art. The "scene category-noise level mapping table" is defined as a mapping from a preset set of scene categories to normalized noise level values ​​(value range [0.00, 1.00]), where the noise level value corresponding to each scene category is obtained by min-max normalization based on the statistical mean of the signal-to-noise ratio of the corresponding scene category samples. The min-max normalization refers to performing (max(X)) normalization on the statistical mean of the signal-to-noise ratio x of each scene category in the scene category set X. x) / (max(X) The min(X) transformation is used to set the normalized noise level value to 0.00 for the scene category with the highest signal-to-noise ratio (quietest) and to 1.00 for the scene category with the lowest signal-to-noise ratio (noisiest).

[0034] Combination Figure 1 as well as Figure 2 Based on the above core concept, the medical device alarm sound recognition method based on scene voiceprint disclosed in this embodiment of the invention specifically includes the following steps: Step S100: Receive the audio segment to be identified. The audio segment to be identified is obtained by segmenting a continuous audio stream from one or more audio acquisition terminals deployed in the medical facility through a fixed-duration sliding window. Overlap between adjacent sliding windows is allowed, and the overlap ratio is determined by system configuration parameters. If the end of the continuous audio stream is less than the preset window duration, it is padded with zero values ​​to the standard length and then used as the audio segment to be identified.

[0035] Step S110: Foreground sound feature extraction and scene voiceprint extraction are performed on the audio segment to be identified to obtain foreground sound features and scene voiceprint. In an embodiment of the present invention, an optional implementation is that the foreground sound features include at least one of spectral energy distribution, zero-crossing rate, temporal envelope, and fundamental frequency estimation.

[0036] In one embodiment of the present invention, an optional implementation is that the scene voiceprint includes a feature vector characterizing the current acoustic scene category.

[0037] In one embodiment of the present invention, an optional implementation method is to extract scene voiceprints by performing short-time Fourier transform and Mel frequency cepstral coefficient analysis on the audio segment to be identified, and then statistically aggregating the obtained spectral features in the time dimension.

[0038] Step S120: Determine whether the overall energy of the audio segment to be identified is lower than the preset silence threshold. If so, it is directly determined as a non-alarm event, and the process returns to step S100 to process the next sliding window; Otherwise, proceed to step S130.

[0039] Step S130: Perform scene classification on the scene voiceprint and calculate the similarity between the scene voiceprint and several pre-constructed scene category prototype vectors. In this embodiment of the invention, an optional implementation is that the pre-constructed scene category prototype vectors include at least three categories: quiet general ward, general ward with people talking, and intensive care ward with multiple devices operating simultaneously.

[0040] Step S140: Determine whether there exists a scene category prototype vector with the highest similarity that exceeds the preset scene classification threshold. If it exists, proceed to step S150; Otherwise, proceed to step S155.

[0041] Step S150: Determine the scene category corresponding to the prototype vector of the current scene category as the scene category to which the current scene voiceprint belongs, and proceed to step S160.

[0042] Step S155: Determine the current scene as an unknown scene and process it according to the preset fallback strategy. In this embodiment of the invention, an optional implementation is that the preset fallback strategy includes either classifying the current audio segment into a preset default noise level and continuing to execute step S160, or determining the current audio segment as a non-alarm event and returning to step S100 to process the next sliding window.

[0043] Step S160: Based on the scene category or unknown scene fallback strategy determined in the previous step, the corresponding noise level estimate is obtained by looking up the pre-built scene category-noise level mapping table.

[0044] In one embodiment of the present invention, an optional implementation is that the scene category-noise level mapping table is defined as a mapping from the scene category set to the normalized noise level value (range 0.00 to 1.00), wherein the normalized noise level value is obtained by normalizing the statistical mean of the signal-to-noise ratio of the corresponding scene category samples.

[0045] Step S170: Determine whether the estimated noise level is less than or equal to the preset noise threshold. If so, proceed to step S180; Otherwise, proceed to step S190.

[0046] Step S180: Match the foreground sound features with each template in the preset waveform template library one by one, and proceed to step S185.

[0047] Step S185: Determine whether there is at least one template whose matching similarity exceeds a preset template matching threshold. If so, proceed to step S200; Otherwise, proceed to step S210.

[0048] Step S210: Determine it as a non-alarm event.

[0049] Step S200: Take the target alarm event corresponding to the template with the highest matching similarity as the judgment result. If multiple templates have the same highest matching similarity, take the target alarm event corresponding to the template with the highest priority according to the preset alarm priority as the judgment result, and proceed to step S220.

[0050] Step S190: Construct a binary feature input by combining the foreground sound features and the scene soundprint, and perform joint matching with each entry in the preset scene-alarm joint mapping table, then proceed to step S195.

[0051] Step S195: Determine whether there exists at least one item whose joint matching similarity exceeds a preset joint matching threshold. If so, proceed to step S200; Otherwise, proceed to step S210.

[0052] In one embodiment of the present invention, an optional implementation is that the identification target of the preset waveform template library and the preset scene-alarm joint mapping table is limited to the set of target alarms pre-registered by the hospital.

[0053] In one embodiment of the present invention, an optional implementation is that the target alarm set includes at least one target alarm tone type.

[0054] In one embodiment of the present invention, an optional implementation method is that the target alarm set is customized by hospital administrators or clinical nurses according to the actual clinical needs of the department. The number of alarm tone types in the target alarm set is one to ten, with a preferred option of one to four types, to accommodate the upper limit of alarm tone types that clinical nurses can effectively identify.

[0055] In one embodiment of the present invention, an optional implementation is to combine... Figure 3 The preset scenario-alarm joint mapping table is constructed using the following method: Acoustic samples were collected from multiple scenarios in actual hospital ward environments, including acoustic samples from quiet general ward environments, acoustic samples from general ward environments with conversations, acoustic samples from intensive care unit environments, and acoustic samples from corridor and nurse station environments. Label the mixed acoustic samples from multiple scenarios, labeling the scenario category to which the sample belongs and whether it contains the target alarm sound and its type (limited to one to ten types). Load a pre-trained large-scale audio representation model as a teacher model, perform joint feature extraction on labeled samples, and obtain a triplet corresponding to each sample. The triplet is represented as (scene voiceprint vector, foreground sound feature vector, alarm type label). In one embodiment of the present invention, an optional implementation is that when the teacher model performs joint feature extraction on the labeled samples: the scene voiceprint vector is extracted from the shallower middle layer of the teacher model; the foreground sound feature vector is extracted from the deeper middle layer of the teacher model or the layer before the event classification head. A lightweight student model is obtained by training triples as the training set and using the knowledge distillation method. In one embodiment of the present invention, an optional implementation is that the total loss during student model training is α. Classification cross-entropy loss + β Distillation loss + γ The scene-assisted classification loss is given by α, β, and γ, which are weights in the system, where α ∈ [0.5, 1.0], β ∈ [0.3, 1.0], and γ ∈ [0.1, 0.5]. The trained student model is compressed using 8-bit integer quantization. The quantized and compressed student model is deployed on the hospital's local edge computing device to serve as an executable implementation of the scene-alarm joint mapping table.

[0056] A specific implementation of the aforementioned preset scenario-alarm joint mapping table includes the following steps: In the actual clinical setting of the target hospital, continuous audio data for at least 30 hours was collected in typical acoustic scenarios such as general wards, intensive care units, surgical preparation areas, corridors, and nurse stations. The audio data was segmented into 1.5-second sliding windows, and annotators labeled each window with its corresponding scenario category, whether it contained the target alarm sound, and its specific type. The target alarm sound types were customized by the hospital's clinical nursing team based on the actual needs of their department. In this embodiment, four types were selected: infusion pump completion alarm, micro-infusion pump obstruction alarm, ventilator low pressure alarm, and monitor heart rate abnormality alarm.

[0057] For labeled window samples, a pre-trained large-scale audio representation model is used as the teacher model for feature extraction. The teacher model can be any of the publicly available audio pre-trained representation models in the field, such as the wav2vec 2.0 model, the HuberT model, the CLAP (Contrastive Language-Audio Pretraining) model, or the PANNs (Pretrained Audio Neural Networks) model or their variants. Scene sound vectors s (typically 128 to 512 dimensions) representing the overall acoustic scene features are extracted from the output of the shallower intermediate layers of the teacher model. Foreground sound feature vectors f (typically 256 to 1024 dimensions) representing the foreground sound event features are extracted from the deeper intermediate layers of the teacher model or the layer before the event classification head. These are combined with manually labeled alarm type labels y to form a training sample set of (s, f, y) triples.

[0058] Using the aforementioned triplet training sample set as training data, a lightweight student model is trained. The student model structure consists of a convolutional neural network with no more than 100 million parameters (preferably no more than 10 million in this embodiment), plus an attention module, such as a MobileNet backbone plus a lightweight self-attention head. The training objective is to maximize the joint prediction accuracy for y given inputs (s, f). The training loss function consists of three weighted parts: L total = α · L CE (y pred , y) + β · L KD (h student , h teacher ) + γ · L scene (s pred,s label ) In the formula, L CE (y pred , y) represents the classification cross-entropy loss for alarm type labels; L KD (h student , h teacher ) represents the distillation loss between the intermediate representation of the student model and the corresponding layer representation of the teacher model (mean squared error or KL divergence can be used); L scene (s pred , s label The loss is the auxiliary classification loss for scene categories; α, β, and γ are weight coefficients, with recommended values ​​in the range of α ∈ [0.5, 1.0], β ∈ [0.3, 1.0], and γ ∈ [0.1, 0.5]. After training, the student model is compressed using 8-bit integer quantization and deployed to the hospital's local edge computing devices (e.g., ARM-based local inference boxes or hospital intranet servers).

[0059] The student model constitutes the executable implementation of the preset scene-alarm joint mapping table; its equivalent representation is a multidimensional lookup table of (scene soundprint category × foreground sound feature clustering × alarm type label).

[0060] Step S220: For results that are determined to be valid alarm events, output the corresponding alarm type label, and convert the alarm type label into voice prompts or text overlay information and push it to the corresponding clinical terminal.

[0061] In one embodiment of the present invention, an optional implementation further includes signal enhancement processing of the foreground sound features and / or suppression processing of the background acoustic components represented by the scene soundprint, wherein the activation conditions for signal enhancement processing and background suppression processing are determined by the category to which the scene soundprint belongs.

[0062] In this embodiment, a preset noise threshold θ is used. n = 0.50, with the preset middle interval being (0.30, 0.50).

[0063] When the noise level n(s(x)) corresponding to the category of the scene voiceprint is ≤0.30 (typical scenario: quiet ordinary ward at night), DSP waveform template matching is directly performed without additional signal enhancement or background suppression processing.

[0064] When the noise level n(s(x)) corresponding to the category of the scene's voiceprint is located in the preset middle range (0.30, 0.50] (typical scenarios: ordinary wards where people talk during the day, surgical preparation areas), it is still within the noise threshold θ nThe determination path is the DSP waveform template matching path; however, before performing template matching, signal enhancement processing is performed on the foreground sound features. The signal enhancement processing improves the matching margin by performing energy normalization and main peak gain on the foreground sound features based on the preset alarm audio spectrum peak template.

[0065] When the noise level n(s(x)) corresponding to the category of the scene's voiceprint is greater than θ n When the value is 0.50 (typical scenario: intensive care unit, corridor and nurse station), the determination path is the scene-alarm joint mapping table matching path; before performing joint matching, background suppression processing is performed on the scene voiceprint. The background suppression processing is used to reduce noise interference in the student model input by subtracting the steady-state background spectrum components that are unrelated to the peak position of the target alarm audio spectrum from the scene voiceprint.

[0066] Accordingly, signal enhancement processing is only enabled in the medium noise sub-interval at the front end of the DSP path, and background suppression processing is only enabled at the front end of the joint mapping path. The two types of processing branches do not overlap in the noise level interval, thus avoiding logical conflicts.

[0067] In one embodiment of the present invention, an optional implementation is as follows: Under three scenarios—initial deployment in a new hospital, changes in scenario category, and audio clips that cannot be categorized into any known scenario—the system behavior is designed as follows: Cold start phase: When the identification system corresponding to this method is deployed for the first time in the target hospital, the preset scene-alarm joint mapping table and scene category-noise level mapping table are both based on a general default mapping table constructed from samples from multiple source hospitals as initial values. The general default mapping table covers the default values ​​of the aforementioned scene categories in this embodiment (quiet general ward, general ward with people talking, intensive care unit ward with multiple devices operating simultaneously, surgical preparation area, corridor and nurse station). In the cold start state, the system directly supports the complete judgment process from steps S100 to S220 using the general default mapping table.

[0068] Local calibration phase: After completing local acoustic sample collection and annotation at the target hospital for a duration of no less than a predefined period (e.g., 30 hours), the teacher model feature extraction and knowledge distillation process described in Example 3 is executed. The resulting localized student model replaces the joint mapping entries in the general default mapping table, and the noise level values ​​in the scene category-noise level mapping table are updated with the normalized result of the statistical mean of the local sample signal-to-noise ratio. Both replacement and update are atomic operations of the storage module, and the system can continue to operate based on the original mapping table during the replacement process.

[0069] Unknown Scene Coverage: When the similarity of the audio segment to be identified to all preset scene category prototype vectors does not exceed the preset scene classification threshold, the current audio segment is classified into the preset default noise level according to the coverage strategy described in step S155, and step S160 is continued; or the current audio segment is determined to be a non-alarm event and the next sliding window is processed. In this embodiment, the preset default noise level is θ. n Nearby conservative values ​​(e.g., with θ) n (Equivalent value) to avoid erroneously triggering the DSP path.

[0070] Fully silent segment processing: When the overall energy of the audio segment to be identified is lower than the preset silence threshold (for example, the absolute value of all sliding window sample values ​​is lower than one ten-thousandth of the maximum dynamic range of the signal, which is equivalent to f(x) being approximately a zero vector), it is directly determined as a non-alarm event according to the additional rules in step S120, skips the subsequent steps, and returns to step S100 to process the next sliding window.

[0071] The above method will be further explained below with different embodiments in combination with practical application scenarios.

[0072] Example 1: Alarm Sound Recognition in a General Ward Environment In this embodiment, the audio acquisition terminal is a microphone array deployed next to the bedside in the ward, with a sampling rate of 16 kHz, and is connected to the local edge computing device via a wired connection.

[0073] The audio segments to be identified are fed into the processing flow in sliding windows of fixed duration (e.g., 1.5 seconds). For each audio segment x within a window, the following operations are performed: Step S1: Perform pre-emphasis, framing, and windowing preprocessing on the audio segment x. Obtain spectral features through short-time Fourier transform, and calculate the foreground sound feature f(x) and scene sound signature s(x) from the spectral features. f(x) includes: the normalized spectral energy distribution vector, zero-crossing rate, temporal envelope peak value, and estimated fundamental frequency. s(x) includes the feature vector formed by concatenating the mean and variance over the entire window time dimension after Mel-frequency cepstral coefficient analysis of the spectrum. The sliding window overlap ratio is exemplarily set to approximately one-third in this embodiment (e.g., a step size of 1.0 second when the window length is 1.5 seconds). In actual deployment, the value is determined through calibration experiments based on the minimum time interval between adjacent alarm tone events in the continuous audio stream.

[0074] Step S2: Compare the similarity of s(x) with the preset scene category prototype vector set one by one to determine the scene category to which s(x) belongs. Scene categories include, but are not limited to: quiet general ward, general ward with people talking, intensive care unit ward with multiple devices operating simultaneously, surgical preparation area, corridor and nurse station. A corresponding normalized noise level n is preset for each scene category. A preset noise threshold θ is also set. n In this embodiment, the value is 0.50. In actual deployment, the value will be taken near the median of the noise level distribution of the local scene category in the target hospital and determined through cross-validation. The following values ​​for the mapping between scene category and noise level are exemplary values ​​in this embodiment and will be determined by local calibration experiments in the target hospital in actual deployment: quiet general ward corresponds to n=0.12; general ward with people talking corresponds to n=0.35; intensive care unit ward with multiple devices operating simultaneously corresponds to n=0.68; surgical preparation area corresponds to n=0.55; corridor and nurse station corresponds to n=0.72. In this embodiment, s(x) is determined to be in the "quiet general ward" category, corresponding to n(s(x))=0.12.

[0075] Step S3: Since n(s(x)) = 0.12 ≤ θ n =0.50, enabling the digital signal processing waveform template matching path. The normalized cross-correlation similarity of f(x) with each template in the preset waveform template library T={T1, T2, T3, T4} is calculated, where T1 is the infusion pump completion alarm sound template pre-registered by the hospital, T2 is the micro-infusion pump obstruction alarm sound template, T3 is the ventilator low-pressure alarm sound template, and T4 is the monitor heart rate abnormality alarm sound template. The preset template matching threshold is 0.80 in this embodiment, but will be determined by calibration experiments during actual deployment. In this embodiment, the matching similarity reaches a maximum value of 0.92 at position T2 (an exemplary value in this embodiment; the actual recognition result will be based on the local deployment test data of the target hospital). Since this similarity exceeds the preset template matching threshold and other templates do not have the highest similarity, T2 is directly taken as the judgment result.

[0076] Step S4: Determine that the event is the "micro-infusion pump blockage alarm" corresponding to T2, output the alarm type label "micro-infusion pump blockage", and broadcast the label to the nurse's mobile terminal after voice synthesis, and at the same time display it on the bedside display screen in the form of text overlay.

[0077] Example 2: Alarm Sound Recognition in a Noisy Environment in the Intensive Care Unit The hardware configuration in this embodiment is the same as that in Embodiment 1.

[0078] The audio segment to be identified x simultaneously includes the heart rate abnormality alarm sound of the monitor in this bed, the tail sound of the same alarm sound from the monitor in the adjacent bed, the conversation between the nurse and the family member, and the continuous noise of the equipment fan.

[0079] Step S1: Same as in Example 1, obtain f(x) and s(x).

[0080] Step S2: s(x) is classified as "intensive care unit with multiple devices operating simultaneously", corresponding to n(s(x)) = 0.68, which is greater than the preset noise threshold θ. n =0.50.

[0081] Step S3: Since n(s(x))>θ n The scene-alarm joint mapping table is enabled for path matching. (f(x), s(x)) is constructed as a binary feature input and fed into a pre-deployed student model M for forward inference. Internally, the student model uses joint features to match the joint mapping (scene voiceprint, foreground features, alarm type) learned during its training phase. The preset joint matching threshold is 0.75 in this embodiment, but will be determined by calibration experiments in actual deployment. In this embodiment, the student model outputs the position with the highest joint matching similarity corresponding to "monitor heart rate abnormality alarm," with a similarity of 0.87 (an exemplary value in this embodiment; the actual recognition result will be based on the local deployment test data of the target hospital), exceeding the preset joint matching threshold.

[0082] Step S4: Determine the event as "heart rate abnormality alarm on monitor", output the corresponding alarm type label and push it to the responsible nurse's terminal with priority after voice synthesis.

[0083] A second aspect of this invention discloses a medical device alarm sound recognition system based on scene voiceprints, such as... Figure 4 As shown, it includes the following modules: An audio receiving module is used to receive audio segments to be identified; The feature extraction module is used to simultaneously perform foreground sound feature extraction and scene voiceprint extraction on the audio segment to be recognized, and output foreground sound features and scene voiceprints. It further includes a foreground sound feature extraction subunit and a scene voiceprint extraction subunit. The scene classification module is used to calculate the similarity between the scene voiceprint and several pre-built scene category prototype vectors, and to determine the scene category corresponding to the scene category of the scene voiceprint by the scene category prototype vector with the highest similarity that exceeds the preset scene classification threshold. The path arbitration module is used to obtain the current noise level estimate by looking up a table in the preset scene category-noise level mapping table according to the scene category to which the scene soundprint belongs, and select the subsequent determination path based on the comparison result of the estimate and the preset noise threshold. A digital signal processing (DSP) matching module is used to perform template matching between foreground sound features and a preset waveform template library when the path arbitration module selects a DSP path; The joint matching module is used to take the foreground sound features and scene soundprint as binary feature inputs when the path arbitration module selects the joint mapping path, and perform joint matching with the preset scene-alarm joint mapping table. The alarm output module is used to convert the results of valid alarm events into voice prompts or text overlay information and push them to the clinical terminal. The storage module is used to store and maintain a preset waveform template library, a preset scene-alarm joint mapping table, and a preset scene category-noise level mapping table.

[0084] The embodiments of the present invention have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. The scope of the present invention is defined by the appended claims and their equivalents. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the present invention, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. A method for identifying medical device alarm tones based on contextual acoustic fingerprints, the method comprising: Includes the following steps: Step 1: Receive the audio segment to be identified from the audio acquisition terminal deployed in the medical facility. The audio segment to be identified is obtained by segmenting the continuous audio stream through a fixed-duration sliding window. Adjacent sliding windows are allowed to overlap. The last segment with a length less than the preset window duration is filled with zero value to the standard length and then processed. Step 2: If the overall energy of the audio segment to be identified is lower than the preset silence threshold, it is directly determined as a non-alarm event, and the process returns to Step 1 to process the next sliding window. Otherwise, foreground sound feature extraction and scene voiceprint extraction are performed simultaneously on the audio segment to be identified to obtain foreground sound features and scene voiceprint; Step 3: Perform scene classification on the scene voiceprint, calculate the similarity between the scene voiceprint and several pre-constructed scene category prototype vectors, and determine the scene category corresponding to the scene category to which the scene voiceprint belongs by the scene category with the highest similarity and exceeding the preset scene classification threshold. If the similarity of all scene category prototype vectors is lower than the preset scene classification threshold, the current scene is determined to be an unknown scene and processed according to the preset fallback strategy. The preset fallback strategy includes either classifying the current audio segment into the preset default noise level and continuing to execute step 4, or determining the current audio segment as a non-alarm event and returning to step 1 to process the next sliding window. Step 4: Based on the scene category to which the scene voiceprint belongs or the unknown scene fallback strategy, query the pre-built scene category-noise level mapping table to obtain the current noise level estimate. Step 5: Determine the audio segment to be identified according to the following conditions and rules: When the noise level estimate is less than or equal to a preset noise threshold, the foreground sound features are matched with each template in the preset waveform template library. If the matching similarity of at least one template exceeds the preset template matching threshold, the target alarm event corresponding to the template with the highest matching similarity is taken as the judgment result. If multiple templates have the same highest matching similarity, the target alarm event corresponding to the template with the highest priority is taken as the judgment result according to the preset alarm priority. Otherwise, it is determined as a non-alarm event. When the noise level estimate is greater than the preset noise threshold, the foreground sound features and the scene voiceprint are constructed as binary feature inputs and jointly matched with the preset scene-alarm joint mapping table. If the joint matching similarity of at least one entry exceeds the preset joint matching threshold, the target alarm event corresponding to the entry with the highest joint matching similarity is taken as the judgment result; otherwise, it is judged as a non-alarm event. The identification targets of the preset waveform template library and the preset scene-alarm joint mapping table are limited to the target alarm set pre-registered by the hospital, and the target alarm set includes at least one target alarm sound type; Step 6: Output the results of the valid alarm events as alarm type labels, and convert them into voice prompts or text overlay information to push to the clinical terminal.

2. The method of claim 1, wherein the medical device alarm sound is identified based on a scene-hashed acoustic fingerprint. In step 2, the foreground sound features include at least one of the following: normalized spectral energy distribution, zero-crossing rate, temporal envelope peak value, and fundamental frequency estimate; the scene sound signature includes a statistically aggregated feature vector based on Mel frequency cepstral coefficients in the sliding window time dimension.

3. The method for recognizing alarm sounds of medical devices based on scene voiceprints as described in claim 1, characterized in that, In step 4, the scene category-noise level mapping table contains at least two scene categories, wherein at least one scene category corresponds to a noise level higher than the preset noise threshold, and at least one scene category corresponds to a noise level lower than or equal to the preset noise threshold. The noise level estimate for each scene category in the scene category-noise level mapping table is obtained by normalizing the statistical mean of the signal-to-noise ratio of the training samples for the corresponding scene category.

4. The method for recognizing alarm sounds of medical devices based on scene voiceprints as described in claim 1, characterized in that, The method for constructing the preset scenario-alarm joint mapping table includes: Step 501: Collect mixed acoustic samples of multiple scenes in the actual acoustic environment of the hospital ward, and label each sample with the scene category and whether it contains the target alarm sound and its type. The types of target alarm sounds are customized by the hospital clinical nursing team according to the actual needs of the department, and the number of types is limited to one to ten. Step 502: Using a pre-trained audio representation model as the teacher model, extract scene voiceprint vectors and foreground sound feature vectors from the labeled samples to form a triplet training sample set including the scene voiceprint vectors, foreground sound feature vectors, and alarm type labels; wherein, the scene voiceprint vectors are extracted from the output of the shallower intermediate layer of the teacher model, and the foreground sound feature vectors are extracted from the output of the deeper intermediate layer of the teacher model or the layer before the event classification head, so that the scene voiceprint vectors represent the overall statistical properties of the acoustic scene and the foreground sound feature vectors represent the local structural properties of the sound event; Step 503: Use the triplet training sample set to train a lightweight student model using the knowledge distillation method. The training objectives of the lightweight student model include the distillation loss between the output feature representation of the teacher model and the classification cross-entropy loss for alarm type labels. Step 504: After quantization and compression, the trained lightweight student model is deployed on the hospital's local edge computing device as an executable implementation of the scene-alarm joint mapping table, which is called by the scene-based voiceprint-based medical device alarm sound recognition method when the joint mapping path is selected.

5. The method for identifying alarm sounds of medical devices based on scene voiceprints as described in claim 4, characterized in that, In step 501, the collection scenarios for the multi-scenario mixed acoustic samples include at least three scenarios: a quiet general ward, a general ward where people are talking, and an intensive care unit ward where multiple devices are running simultaneously; in step 503, the number of parameters of the lightweight student model does not exceed 100 million and does not exceed one-tenth of the number of parameters of the teacher model.

6. The method for recognizing alarm sounds of medical devices based on scene voiceprints as described in claim 1, characterized in that, The target alarm set includes one to ten alarm tone types. The target alarm set can be customized by hospital administrators or clinical nurses according to the actual clinical needs of the department, and can be reconfigured locally in the hospital through the configuration interface.

7. The method for recognizing alarm sounds of medical devices based on scene voiceprints as described in claim 1, characterized in that, It also includes a step of signal enhancement processing of the foreground sound features, wherein the signal enhancement processing includes energy normalization and main peak gain of the foreground sound features based on a preset alarm audio spectrum peak template; the signal enhancement processing is performed before the template matching step described in step 5, and its activation condition is that the noise level estimate is located in a preset intermediate interval, and the upper limit of the preset intermediate interval does not exceed the preset noise threshold.

8. The method for recognizing alarm sounds of medical devices based on scene voiceprints as described in claim 1, characterized in that, It also includes a step of suppressing the background acoustic components represented by the scene soundprint. The suppression process includes subtracting steady-state background spectrum components that are independent of the peak position of the target alarm audio spectrum from the scene soundprint. The suppression process is performed before the joint matching step described in step 5, and its activation condition is that the noise level estimate is higher than the preset noise threshold.

9. The method for identifying alarm sounds of medical devices based on scene voiceprints as described in claim 1, characterized in that, In step 1, the audio acquisition terminal is a microphone or microphone array deployed next to the bedside in the ward, operating room or corridor. The medical device alarm sound recognition method and the medical device data acquisition system based on the protocol interface run in parallel, and the alarm recognition result and the protocol acquisition result are sent to the hospital clinical information system together.

10. A medical device alarm sound recognition system based on scene voiceprint, characterized in that, include: The audio receiving module is used to receive audio segments to be identified from audio acquisition terminals deployed in medical facilities. The audio segments to be identified are obtained by segmenting a continuous audio stream through a fixed-duration sliding window. Adjacent sliding windows are allowed to overlap. Segments that are less than the preset window duration at the end are filled with zero values ​​to the standard length and then used as the audio segments to be identified. The feature extraction module is used to simultaneously perform foreground sound feature extraction and scene voiceprint extraction on the audio segment to be identified, and output the foreground sound features and scene voiceprint; and when the overall energy of the audio segment to be identified is lower than the preset silence threshold, directly determine that the current audio segment is a non-alarm event and use the next sliding window as a new audio segment to be identified; The scene classification module is used to calculate the similarity between the scene voiceprint and several pre-constructed scene category prototype vectors respectively, and determine the scene category corresponding to the scene category to which the scene voiceprint belongs by the scene category corresponding to the scene category prototype vector with the highest similarity and exceeding the preset scene classification threshold; and determine the current scene as an unknown scene and process it according to the preset fallback strategy when the similarity of all scene category prototype vectors is lower than the preset scene classification threshold. The path arbitration module is used to query a pre-built scene category-noise level mapping table to obtain the current noise level estimate based on the scene category to which the scene voiceprint belongs or the unknown scene fallback strategy, and select the subsequent determination path based on the comparison result of the noise level estimate and the preset noise threshold. The DSP matching module is used to perform template matching between the foreground sound features and each template in the preset waveform template library when the path arbitration module selects a digital signal processing path. If the matching similarity of at least one template exceeds the preset template matching threshold, the target alarm event corresponding to the template with the highest matching similarity is taken as the judgment result. If the matching similarity of multiple templates is tied for the highest, the target alarm event corresponding to the template with the highest priority is taken as the judgment result according to the preset alarm priority. Otherwise, it is determined as a non-alarm event. The joint matching module is used to construct the foreground sound features and the scene voiceprint as binary feature inputs when the path arbitration module selects the joint mapping path, and perform joint matching with the preset scene-alarm joint mapping table. Based on the comparison result of the joint matching similarity and the preset joint matching threshold, it is determined whether it is the target alarm event corresponding to the corresponding entry. The alarm output module is used to output the results of valid alarm events as alarm type labels and convert them into voice prompts or text overlay information to be pushed to the clinical terminal. The storage module is used to store and maintain the preset waveform template library, the preset scene-alarm joint mapping table, and the preset scene category-noise level mapping table. The identification targets of the preset waveform template library and the preset scene-alarm joint mapping table are limited to a target alarm set consisting of one to ten alarm sound types pre-registered by the hospital.

Images